Construction method and application of bone age evaluation model based on label distribution learning
By constructing a bone age assessment model based on label distribution learning and utilizing the normal probability density function and improved loss function, the problems of low efficiency and insufficient robustness of traditional methods are solved, and high-precision bone age assessment is achieved.
Patent Information
- Application Number
- CN202510490467.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-18
- Publication Date
- 2025-10-10
AI Technical Summary
Traditional bone age assessment methods rely on doctors' experience, are inefficient and highly subjective. Existing deep learning methods lack robustness and generalization capabilities when data sets are incomplete and incorrectly labeled, making it difficult to effectively utilize inter-class features and sequentiality of bone age.
A bone age assessment model based on label distribution learning was constructed. The training sample set was divided into multiple age groups by clustering. The sub-model and convolutional network were used to integrate the probability distribution of bone age labels. The normal probability density function and improved loss function were introduced, and training was performed in combination with time series combination labels.
It improves the accuracy and robustness of bone age assessment, reduces the misjudgment of adjacent bone ages, enhances the model's adaptability to missing data sets and labeling errors, and improves the accuracy of bone age prediction and the stability of the model.
Smart Images
Figure CN120765522A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of medical image processing and artificial intelligence, and more specifically, relates to a method for constructing a bone age assessment model based on label distribution learning and its application. Background Art
[0002] Bone age (BA) refers to the developmental age of bones and is an important indicator for assessing the growth and development of children and adolescents. Traditionally, BA assessment relies primarily on a physician's visual inspection of hand bone X-ray images, comparing them with standard atlases (such as the Greulich-Pyle atlas). However, this method has drawbacks such as high subjectivity, low efficiency, and susceptibility to physician experience. With the advancement of medical imaging technology, automated BA assessment methods based on image processing and artificial intelligence have gradually become a research hotspot.
[0003] Defects and shortcomings of existing technology:
[0004] Traditional bone age assessment methods rely primarily on physician experience and standardized atlases. This approach is not only inefficient but also highly subjective, with significant variability in assessments between physicians. Furthermore, this approach fails to fully leverage the advantages of modern medical imaging technology and artificial intelligence algorithms, making automated bone age assessment difficult.
[0005] With the development of deep learning technology, bone age assessment methods based on convolutional neural networks (CNNs) have gradually become a research hotspot. These methods typically treat bone age assessment as a classification or regression task. However, the small differences between bone age classes and their strong sequential nature make it difficult for traditional classification or regression methods to effectively utilize inter-class features, leading to the misclassification of adjacent bone ages. For example, labels represented using one-hot encoding or regression values cannot reflect the correlation and sequential nature between bone age classes.
[0006] Existing deep learning-based bone age assessment methods perform poorly when dealing with incomplete and mislabeled datasets. These methods often require high dataset quality and are unable to effectively handle noise and outliers. Consequently, these methods lack robustness and generalization capabilities in practical applications, making them difficult to meet clinical needs. Summary of the Invention
[0007] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides a method for constructing and applying a bone age assessment model based on label distribution learning, the purpose of which is to improve the accuracy of bone age assessment.
[0008] To achieve the above objectives, according to one aspect of the present invention, a method for constructing a bone age assessment model based on label distribution learning is provided, comprising:
[0009] Construct a training sample set containing hand X-ray images and their corresponding true bone age labels;
[0010] A bone age assessment model was constructed, comprising a sub-model A, K sub-models B, a selection unit, and a convolutional network C. The value of K was determined by dividing the entire training sample set into K intervals using a clustering algorithm or prior knowledge to meet the bone age assessment needs of different age groups. The K intervals corresponded to different age groups.
[0011] Use the entire training sample set to train sub-model A, and use the training sample set corresponding to each interval to train sub-model B corresponding to the age group;
[0012] Among them, the sub-model A is used to predict the bone age label probability distribution according to each input image, and obtain the predicted bone age label probability distribution corresponding to the input image of the sub-model A; the selection unit is used to bring the predicted bone age label probability distribution output by the sub-model A into the pre-constructed normal probability density function centered on the true bone age of the input image, and calculate the corresponding normal distribution probability value; and select the sub-model B of the age group corresponding to the age of the input image; the sub-model B is used to predict the bone age label probability distribution again according to the input image, and obtain the predicted bone age label probability distribution of the sub-model B corresponding to the age group of the input image; the convolutional network C is used to integrate the predicted bone age label probability distributions from the sub-model A and the selected sub-model B, and generate the final predicted bone age label probability distribution, completing the construction of the bone age assessment model based on label distribution learning.
[0013] Furthermore, the training sample set is obtained through data preprocessing, wherein the data preprocessing includes image cropping, normalization and data enhancement operations.
[0014] Furthermore, the normal probability density function centered on the true bone age of each input image is:
[0015]
[0016] Where A gt is the true bone age of the input image, C is the total number of bone age categories, x represents the bone age category, and σ is the standard deviation of the control label distribution range, which serves as a hyperparameter.
[0017] Furthermore, the loss functions used in training each sub-model include:
[0018] The loss function that minimizes the mean absolute error (MAE) between the network's predicted output bone age value and the actual bone age value is And the loss function that measures the difference between the predicted bone age label probability distribution and the true bone age label probability distribution
[0019] the loss function and the loss function are added as the total loss function
[0020] Further, the calculation formula of the mean absolute error MAE is:
[0021]
[0022] The calculation formula of the KL divergence is:
[0023]
[0024] In the formula, n represents the total number of training samples, y i represents the true bone age label probability distribution of the i-th training sample, p i represents the predicted bone age label probability distribution of the i-th training sample by the sub-model; the superscript T represents transposition.
[0025] Further, in the training process of each sub-model, a time sequence combined label is introduced to correct the wrong predicted bone age label corresponding to the sub-model; wherein the calculation formula of the time sequence combined label is:
[0026] s t = αs t-1 + (1-α)p t
[0027]
[0028] Wherein, s t and s t-1 respectively represent the output prediction combination of the sub-model at the t-th and (t-1)-th epochs, p t represents the output prediction distribution at the t-th epoch, represents the current time sequence combined label at the t-th epoch.
[0029] Further, as the training proceeds, the weight of the time sequence combined label in the loss function is gradually increased, and the specific weight adjustment formula is:
[0030]
[0031] Wherein, T(t) represents the weight of the loss function corresponding to the time sequence combined label, T MAX represents the maximum value of T(t), and E represents the total number of epochs.
[0032] According to another aspect of the present invention, a bone age assessment method based on label distribution learning is provided, comprising:
[0033] The currently collected test set is input into the bone age assessment model to obtain a bone age assessment value; wherein, the bone age assessment model is a bone age assessment model constructed by the construction method of a bone age assessment model based on label distribution learning as described above.
[0034] According to another aspect of the present invention, a bone age assessment system based on label distribution learning is provided, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the bone age assessment method as described above when executing the computer program.
[0035] According to another aspect of the present invention, a machine-readable storage medium is provided, which stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the method for constructing a bone age assessment model as described in any one of the above items and / or the bone age assessment method as described above.
[0036] In general, the above technical solutions conceived by the present invention have the following beneficial effects compared with the prior art:
[0037] 1. The present invention provides a method for constructing a bone age assessment model based on label distribution learning, wherein the entire training sample set is used to train sub-model A, and the training sample set corresponding to each age interval is used to train sub-model B of the corresponding age group, and the selection unit is used to bring the predicted bone age label probability distribution output by sub-model A into a pre-constructed normal probability density function centered on the true bone age of the input image to calculate the corresponding normal distribution probability value; and select sub-model B of the age group corresponding to the age of the input image; sub-model B is used to predict the bone age label probability distribution again based on the input image to obtain the predicted bone age label probability distribution of sub-model B corresponding to the age group of the input image; convolutional network C is used to integrate the predicted bone age label probability distribution from sub-model A and the selected sub-model B, and generate the final predicted bone age label probability distribution, by combining models with different training fineness, improving the accuracy of bone age prediction; and by converting the traditional single numerical bone age label into a label distribution that conforms to the law of normal distribution, the model can better capture and utilize the correlation and sequentiality between bone age classes. This transformation of the label distribution allows the model to take into account the similarities of adjacent bone age categories during feature extraction and classification, thereby reducing the problem of misclassification of adjacent bone ages.
[0038] 2. The present invention provides a method for constructing a bone age assessment model based on label distribution learning. The loss function introduced during training combines the mean absolute error (MAE) and Kullback-Leibler (KL) divergence, which not only minimizes the deviation between the predicted bone age and the true bone age, but also optimizes the difference between the predicted label distribution and the true label distribution, further improving the accuracy of bone age assessment.
[0039] 3. The construction method of the bone age assessment model based on label distribution learning provided by the present invention introduces a time series combination label and an improved loss function so that the model can better handle the missing categories and labeling errors in the data set, and enhances the adaptability to the data set. Especially in the case where the bone age data set sample is incomplete or the labeling is wrong, the inventive method can effectively reduce the impact of these problems and improve the robustness of the model. In addition, by combining the prediction output of different training cycles, the robustness of the model to error labels is improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 A flowchart of a method for constructing a bone age assessment model based on label distribution learning provided by an embodiment of the present invention;
[0041] Figure 2 Schematic diagram of the training and evaluation process of the bone age assessment model based on label distribution learning provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0042] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
[0043] Example 1
[0044] A method for constructing a bone age assessment model based on label distribution learning, such as Figure 1 Shown, including:
[0045] Construct a training sample set containing hand X-ray images and their corresponding true bone age labels;
[0046] A bone age assessment model was constructed, comprising a sub-model A, K sub-models B, a selection unit, and a convolutional network C. The value of K was determined by dividing the entire training sample set into K intervals using a clustering algorithm or prior knowledge to meet the bone age assessment needs of different age groups. The K intervals corresponded to different age groups.
[0047] The entire training sample set is used to train sub-model A, and the training sample set corresponding to each interval is used to train sub-model B of the corresponding age group; wherein, sub-model A is used to predict the bone age label probability distribution according to each input image, and obtain the predicted bone age label probability distribution corresponding to the input image of the sub-model A; the selection unit is used to bring the predicted bone age label probability distribution output by sub-model A into a pre-constructed normal probability density function centered on the true bone age of the input image, and calculate the corresponding normal distribution probability value; and select the sub-model B of the age group corresponding to the age of the input image; the sub-model B is used to predict the bone age label probability distribution again according to the input image, and obtain the predicted bone age label probability distribution of the sub-model B corresponding to the age group of the input image; the convolutional network C is used to integrate the predicted bone age label probability distributions from sub-model A and the selected sub-model B, and generate the final predicted bone age label probability distribution, completing the construction of the bone age assessment model based on label distribution learning.
[0048] Specifically, according to the characteristics of the bone age data within each bone age label category, the parameters of the normal distribution are set, where the standard deviation is used to control the range of the bone age label distribution. The closer the distance to the bone age label category, the greater the output probability of the corresponding bone age category; the farther the distance from the bone age label category, the smaller the output probability of the corresponding bone age category. Therefore, when quantifying the category labels of the image dataset, the bone age label should be expressed as a complete and ordered set y = y0, y1, ..., y C-1 , where C is the total number of categories, and since the same step length is used when quantizing y, it is logical to assume that the predicted bone age label distribution y obeys the true bone age A gt The normal probability density function G(x) centered at is:
[0049]
[0050] Among them, A gt is the true bone age, C is the total number of bone age categories, and σ is the standard deviation of the control label distribution range, which is a hyperparameter. By adjusting σ, the fuzziness of the label distribution can be controlled. From this, the rationality of the inferred σ range determines the validity of the label distribution.
[0051] Convert the single-value bone age label to a label distribution that conforms to the normal distribution law. For each bone age category c, calculate its corresponding normal distribution probability value Form a label distribution.
[0052] Preferably, the loss function of the bone age assessment model includes: mean absolute error (MAE) and Kullback-Leibler (KL) divergence, wherein the MAE is used to minimize the deviation between the predicted bone age and the true bone age, and the KL divergence is used to measure the difference between the predicted bone age label probability distribution and the true bone age label probability distribution.
[0053] Specifically, given a sample with a re-established label distribution, the label distribution is y and the true bone age is A gt =argmax(y), assuming is the activation value of the last layer of the bone age assessment network, which is converted into a probability distribution p with a sum of 1 using the softmax function, that is, the predicted distribution of the network output.
[0054] Furthermore, it is expected that the network will be trained through the loss function to minimize the MAE between the network predicted output bone age value and the actual bone age value, so there is a loss function
[0055]
[0056] At the same time, using only MAE cannot minimize the difference between two distributions, so Kullback-Leibler (KL divergence) is introduced to measure the degree of difference between two distributions. When the KL divergence is larger, the difference between the two distributions is greater and the similarity is smaller; when the KL divergence value is smaller, the difference between the two distributions is smaller and the similarity is greater; when the two distributions are the same, the KL divergence value is 0.
[0057] So there is a loss function
[0058]
[0059] Furthermore, the loss function is
[0060]
[0061] The above-mentioned bone age assessment model is trained using a pre-collected training set.
[0062] The training sample set includes hand bone images and their corresponding single numerical true bone age labels. During the training process, the label distribution and loss function are used to guide model learning, enabling the model to fully explore the characteristics of adjacent bone age classes and improve robustness to erroneous labels.
[0063] Specifically, a training set consisting of hand bone images and their corresponding true bone age labels is constructed. This training set is obtained through data preprocessing, which includes image cropping, normalization, and data augmentation. The input hand bone X-ray images are fed into the model training process. The entire dataset is partitioned into K intervals using a clustering algorithm or prior knowledge to accommodate bone age assessment needs across different age groups. These intervals correspond to different age groups, such as 3-12 years, 12-17 years, and 17-20 years.
[0064] Furthermore, ifFigure 2 As shown in the figure, the training set is used to train the bone age assessment model. For each age group, the dataset is further divided and trained separately to obtain a specialized sub-model. For example, in the 3-12 age range, model B1 is trained; in the 12-17 age range, model B2 is trained; and in the 17-20 age range, model B3 is trained. This partitioned training method allows the model to focus more on feature learning for specific age groups, thereby improving the accuracy and efficiency of bone age assessment. During the training process, the label distribution and loss function are used to guide model learning, enabling the model to fully explore the features of adjacent bone age classes and improve robustness to incorrect labels.
[0065] Furthermore, considering the particularity of the bone age dataset, the labeled labels may not be real labels, and when the number of such samples is small, During training, the adjacent classes of the class with the highest probability value will be paid attention to, and the adjacent classes will contribute to the correct prediction of the middle class, and the probability value of the potential correct class will gradually increase. If the incorrect predicted bone age label corresponding to the sub-model is further corrected, the accuracy of bone age assessment will be further improved. For this purpose, a time series combination label is proposed.
[0066] s t =αs t-1 +(1-α)p t
[0067]
[0068] Among them, s t and s t-1 They represent the network output prediction combination at the tth and t-1th epochs respectively. Similarly, p t Indicates the output prediction distribution at the tth epoch, represents the current time series combination label at the tth epoch. As training progresses, the more recent prediction distribution will be more reliable than the earlier prediction, so the hyperparameter coefficient α∈[0,1] is added to control p t In addition, in order to correct s t Deviation at system startup, s t Divide by 1-α t , so that the time series combination label of the next epoch can be obtained.
[0069] Furthermore, as the network training progresses, the network's ability to eliminate ambiguous labels becomes stronger and stronger. At this time, the temporal combination label It will be more effective than the label distribution y and can more effectively guide the training of the network, so T(t) is added to control The weight of , the total loss function is
[0070]
[0071] Among them, T MAX is the maximum value of T(t), and E is the total number of epochs, so that T(t) can be guaranteed to increase gradually from 0. It should be noted that when the network has a multi-branch progressive structure, the loss function It is usually broken down into different parts to guide the training of different branches.
[0072] Furthermore, the trained bone age assessment model is evaluated using a validation set. The accuracy and robustness of the model are calculated to verify the model's performance. After model training is completed, for a given input image, preliminary feature extraction and classification are first performed using model A. Then, based on the output of model A, the corresponding sub-model (such as model B1, B2, or B3) is selected for further prediction of the bone age label probability distribution. Each sub-model outputs a predicted bone age label probability distribution corresponding to its training age group.
[0073] Finally, these predicted bone age label probability distributions are fed into a network layer that integrates the predicted bone age label probability distributions from different sub-models to generate the final predicted bone age label probability distribution. If the predicted bone age A is equal to 16 years old, the result is directly output; otherwise, the integrated probability distribution is normalized through a softmax layer to obtain the final bone age assessment result.
[0074] Example 2
[0075] A bone age assessment method based on label distribution learning, comprising:
[0076] The currently collected test set is input into the bone age assessment model to obtain a bone age assessment value; wherein, the bone age assessment model is a bone age assessment model constructed by a method for constructing a bone age assessment model based on label distribution learning as described in Example 1.
[0077] The relevant technical solutions are the same as those in Example 1 and will not be described again here.
[0078] Example 3
[0079] A bone age assessment system based on label distribution learning includes: a memory and a processor, wherein the memory stores a computer program, and the processor executes the bone age assessment method as described in Example 2 when executing the computer program.
[0080] The system can be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The processor can be a central processing unit (CPU), or other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store computer programs and / or modules, and the processor can run or execute computer programs and / or modules stored in the memory, as well as call data stored in the memory, to realize various functions of the electronic device.
[0081] The relevant technical solutions are the same as above and will not be repeated here.
[0082] Example 4
[0083] A machine-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the bone age assessment model construction method and / or bone age assessment method of the above-mentioned embodiment.
[0084] Specifically, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.
[0085] The relevant technical solutions are the same as above and will not be repeated here.
[0086] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for constructing a bone age assessment model based on label distribution learning, characterized in that: include: Construct a training sample set containing hand X-ray images and their corresponding true bone age labels; A bone age assessment model was constructed, comprising a sub-model A, K sub-models B, a selection unit, and a convolutional network C. The value of K was determined by dividing the entire training sample set into K intervals using a clustering algorithm or prior knowledge to meet the bone age assessment needs of different age groups. The K intervals corresponded to different age groups. Use the entire training sample set to train sub-model A, and use the training sample set corresponding to each interval to train sub-model B corresponding to the age group; Among them, the sub-model A is used to predict the bone age label probability distribution according to each input image, and obtain the predicted bone age label probability distribution corresponding to the input image of the sub-model A; the selection unit is used to bring the predicted bone age label probability distribution output by the sub-model A into the pre-constructed normal probability density function centered on the true bone age of the input image, and calculate the corresponding normal distribution probability value; and select the sub-model B of the age group corresponding to the age of the input image; the sub-model B is used to predict the bone age label probability distribution again according to the input image, and obtain the predicted bone age label probability distribution of the sub-model B corresponding to the age group of the input image; the convolutional network C is used to integrate the predicted bone age label probability distributions from the sub-model A and the selected sub-model B, and generate the final predicted bone age label probability distribution, completing the construction of the bone age assessment model based on label distribution learning.
2. The construction method according to claim 1, wherein The training sample set is obtained through data preprocessing, wherein the data preprocessing includes image cropping, normalization and data enhancement operations.
3. The construction method according to claim 1, wherein The normal probability density function centered on the true bone age of each input image is: Where A gt is the true bone age of the input image, C is the total number of bone age categories, x represents the bone age category, and σ is the standard deviation of the control label distribution range, which serves as a hyperparameter.
4. The construction method according to claim 1, wherein The loss functions used in training each sub-model include: The loss function that minimizes the mean absolute error (MAE) between the network's predicted output bone age value and the actual bone age value is And the loss function that measures the difference between the predicted bone age label probability distribution and the true bone age label probability distribution The loss function And the loss function The sum of , as the total loss function when training the sub-model 5. The construction method according to claim 4, wherein: The calculation formula of the mean absolute error MAE is: The calculation formula of the KL divergence is: Where n represents the total number of training samples, y i represents the true bone age label probability distribution of the i-th training sample, p i Represents the probability distribution of the predicted bone age label of the i-th training sample by the sub-model; the superscript T indicates transposition.
6. The construction method according to claim 1, wherein: During the training process of each sub-model, the temporal combination label is introduced It is used to correct the erroneous predicted bone age label corresponding to the sub-model, wherein the calculation formula of the time series combination label is: s t =αs t-1 +(1-α)p t Among them, s t and s t-1 Respectively represent the output prediction combination of the sub-model at the tth and t-1th epochs, p t Indicates the output prediction distribution at the tth epoch, Indicates the current time series combination label at the tth epoch.
7. The construction method according to claim 6, wherein: As training progresses, the weight of the time series combination label in the loss function is gradually increased. The specific weight adjustment formula is: Among them, T(t) represents the loss function corresponding to the time series combination label The weight, T MAX represents the maximum value of T(t), and E represents the total number of epochs.
8. A bone age assessment method based on label distribution learning, comprising: The currently collected test set is input into the bone age assessment model to obtain a bone age assessment value; wherein, the bone age assessment model is a bone age assessment model constructed by a method for constructing a bone age assessment model based on label distribution learning as described in any one of claims 1 to 7.
9. A bone age assessment system based on label distribution learning, characterized in that: include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the bone age assessment method according to claim 8 when executing the computer program.
10. A machine-readable storage medium, characterized in that The machine-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by the processor, the machine-executable instructions prompt the processor to implement the method for constructing a bone age assessment model as described in any one of claims 1 to 7 and / or the bone age assessment method as described in claim 8.